Document 11225460

advertisement
Penn Institute for Economic Research
Department of Economics
University of Pennsylvania
3718 Locust Walk
Philadelphia, PA 19104-6297
pier@econ.upenn.edu
http://economics.sas.upenn.edu/pier
PIER Working Paper 15-015
“Skewed Noise”
by
David Dillenberger and Uzi Segal
http://ssrn.com/abstract=2580407
Skewed Noise∗
David Dillenberger† Uzi Segal‡
March 18, 2015
We study the attitude of decision makers to skewed noise. For a binary lottery
that yields the better outcome with probability p, we identify noise around p
with a compound lottery that induces a distribution over the exact value of
the probability and has an average value p. We propose and characterize a
new notion of skewed distributions, and use a recursive non-expected utility
model to provide conditions under which rejection of symmetric noise implies
rejection of skewed to the left noise as well. We demonstrate that rejection
of these types of noises does not preclude acceptance of some skewed to the
right noise, in agreement with recent experimental evidence. We apply the
model to study random allocation problems (one-sided matching) and show
that it can predict systematic preference for one allocation mechanism over
the other, even if the two agree on the overall probability distribution over
assignments. The model can also be used to address the phenomenon of
ambiguity seeking in the context of decision making under uncertainty.
JEL Classification number: D81, C78
Keywords: Skewed distributions, recursive non-expected utility, ambiguity
seeking, one-sided matching.
∗
We thank Faruk Gul, Yoram Halevy, Efe Ok, and Rakesh Vohra, as well as the coeditor and four anonymous referees, for their comments and help. We thank Francesca
Toscano and Zhu Zhu of Boston College for helpful research assistance.
†
Department of Economics, University of Pennsylvania (ddill@sas.upenn.edu).
‡
Department of Economics, Boston College (segalu@bc.edu) and Warwick Business
School.
1
Introduction
Standard models of decision making under risk assume that individuals obey
the reduction of compound lotteries axiom, according to which a decision
maker is indifferent between any multi-stage lottery and the simple lottery
that induces the same probability distribution over final outcomes. Experimental and empirical evidence suggest, however, that this axiom is often
violated (see, among others, Kahneman and Tversky [25], Bernasconi and
Loomes [5], Conlisk [13], Harrison, Martinez-Correa, and Swarthout [22],
and Abdellaoui, Klibanoff, and Placido [1]). Individuals may have preferences over gambles with identical probability distributions over final outcomes if they differ in the form of the timing of resolution of uncertainty.
Also, individuals may view ambiguity as lotteries over the true values of
the probabilities and may simply like or dislike not knowing their exact values. For example, subjects are typically not indifferent between betting on
a known probability p and betting on a known distribution over the value of
that probability even when the mean probability of the distribution is p.
Halevy [21] and recently Miao and Zhong [32], for example, consider preferences over two-stage lotteries and demonstrate that individuals are averse
to the introduction of symmetric noise, that is, symmetric mean-preserving
spread into the first-stage lottery. One rationale for this kind of behavior is
that the realizations in a symmetric noise cancel out each other and simply
create an undesired confusion in evaluation. On the other hand, asymmetric
noises, and in particular positively skewed ones, may be desirable. Boiney [6]
conducted an experiment in which decision makers had to choose one of three
investment plans. In all three prospects, the overall probability of success
(which results in a prize x = $200) is p = 0.2, and with the remaining probability the investment fails and the decision maker receives x = $0. Option A
represents an investment plan in which the decision maker is confident about
the probability of success. In B and C, on the other hand, the probability
of success is uncertain. Prospect B (resp., C) represents a negatively (positively) skewed distribution around p in which it is very likely that the true
probability slightly exceeds (falls below) p but it is also possible, albeit unlikely, that the true probability is much lower (higher). Boney’s main finding
is that decision makers are not indifferent between the three prospects and
that most prefer C to A and A to B. Moreover, these preferences are robust
to different values of x, x, and p.
In Boney’s experiment, the underlying probability of success p was the
2
same in all three options. In a recent experiment, which we discuss in more
detail in Section 3, Abdellaoui, Klibanoff, and Placido [1] found strong evidence that aversion to compound risk (i.e., noise) is an increasing function
of p. In particular, their results are consistent with a greater aversion to
negatively skewed noise around high probabilities than to positively skewed
noise around small probabilities.
In this paper we propose a model that can accommodate the behavioral
patterns discussed above. For a binary lottery that yields the better outcome
with probability p, we identify noise around p with a two-stage lottery that
induces a distribution over the exact value of the probability and has an
average value p. We introduce and characterize a new notion of skewness, and
use a version of Segal’s [38] recursive non-expected utility model to outline
conditions under which a decision maker who always rejects symmetric noise
will also reject any negatively skewed noise (for instance, will prefer option
A to B in the example above) but may seek some positively skewed noise.
We suggest two applications. First, we apply our model to study a simple
variant of the house allocation problem (or one-sided matching), where the
goal is to look for a systematic way of assigning a set of indivisible objects to
a group of individuals having preferences over these objects. We demonstrate
that different mechanisms, which agree on the overall probability distribution
over assignments and hence are being treated equivalently in the standard
model, induce different compound lotteries. Comparing two familiar mechanisms, versions of the random serial dictator and of the random top cycle,
our model predicts systematic preference for the latter over the former for a
large set of parameters, while permitting the opposite preferences when one
type of the goods is scarce, but almost everyone prefers it over the alternative
type (see Section 4 for details).
Our second application shows that our model can be used to address the
recently documented phenomenon of ambiguity seeking in the context of decision making under uncertainty. The recursive model we study here was
first suggested by Segal [37] as a formal way to analyze attitudes towards
ambiguity. Under this interpretation, ambiguity is identified as a two-stage
lottery, where the first stage captures the decision makers subjective uncertainty about the true probability distribution over the states of the world,
and the second stage determines the probability of each outcome, conditional
on the probability distribution that has been realized. Our model permits
the co-existence of aversion to symmetric ambiguity (as in Ellsberg’s famous
paradox) and ambiguity seeking, especially in situations where the decision
3
maker anticipates a bad outcome, yet believes that there is a small chance
that things are not as bad as they seem. In this case, he might not want to
know the exact values of the probabilities.
The fact that the recursive evaluation of two-stage lotteries in Segal’s
model is done using non-expected utility functionals is key to our analysis.
It is easy to see that if the decision maker uses the same expected utility
functional in each stage he will be indifferent to noise. In Section 6 we
further show that a version of the recursive model in which the two stages are
evaluated using different expected utility functionals (Kreps and Porteus [26],
Klibanoff, Marinacci, and Mukerji [25]) cannot accommodate the co-existence
of rejecting all symmetric noise while still accepting some positively skewed
noise.
In this paper, we confine our attention to the analysis of attitudes to
noise related to the probability p in the binary prospect which pays x with
probability p and x otherwise, where x > x. In reality the decision maker
may face lotteries with many outcomes and the probabilities of receiving
each of them may be uncertain. We prefer to deal only with binary lotteries
since when there are many outcomes their probabilities depend on each other
and therefore skewed noise over the probability of one event may affect noises
over other probabilities in too many ways. This complication is avoided when
there are only two outcomes — whatever the decision maker believes about
the probability of receiving x completely determines his beliefs regarding the
probability of receiving x. Note that while the underlying lottery is binary,
the noise itself (that is, the distribution over the value of p) may have many
possible values or may even be continuous.
The rest of the paper is organized as follows: Section 2 describes the
analytical framework and introduces notations and definitions that will be
used in our main analysis. Section 3 studies attitudes towards asymmetric
noises and states our main behavioral result. Section 4 and Section 5 are
devoted to applications. Section 6 comments on the relationship of our paper
to other models. All proofs are relegated to the appendix.
4
2
The model
2.1
Preferences
Fix two monetary outcomes x > x. The underlying lottery we consider is
the binary prospect (x, p; x, 1 − p), which pays x with probability p and x
otherwise. We identify this lottery with the number p ∈ [0, 1] and analyze
noise around p as a two-stage lottery, denoted by hp1 , q1 ; ...; pn , qn i, that yields
with probability qi the lottery (x, pi ; x, 1 − pi ), i = 1, 2, ..., n, and satisfies
P
i pi qi = p. Let
P
L2 = {hp1 , q1 ; ...; pn , qn i : pi , qi ∈ [0, 1] , i = 1, 2, ..., n, and i qi = 1} .
Let be a complete and transitive preference relation over L2 , which is
represented by U : L2 → ℜ. Throughout the paper we confine our attention
to preferences that admit the following representation:
U (hp1 , q1 ; ...; pn , qn i) = V (cp1 , q1 ; ...; cpn , qn )
(1)
where V is a functional over simple (finite support) one-stage lotteries over
the interval [x, x] and c is a certainty equivalent function (not necessarily the
one obtained from V ).1 According to this model, the decision maker evaluates
a two-stage lottery hp1 , q1 ; ...; pn , qn i recursively. He first replaces each of the
second-stage lotteries with its certainty equivalent, cpi . This results in a
simple, one-stage lottery over the certainty equivalents, (cp1 , q1 ; ...; cpn , qn ),
which he then evaluates using the functional V .2 We assume throughout
that V is monotonic with respect to first-order stochastic dominance and
continuous with respect to the weak topology.
There are several reasons that lead us to study this special case of U .
First, it explicitly captures the sequentiality aspect of two-stage lotteries,
by distinguishing between the evaluations made in each stage (V and c in
the first and second stage, respectively). Second, it allows us to state our
results using familiar and easy to interpret conditions that are imposed on
1
The function c: [0, 1] → ℜ is a certainty equivalent function if for some W over onestage lotteries, W (cp , 1) = W (x, p; x, 1 − p).
2
The functional V thus represents some underlying complete and transitive binary
relation over simple lotteries, which is used in the first stage to evaluate lotteries over the
certainty equivalents of the second stage. To avoid confusion with the main preferences
over L2 , we will impose all the assumptions in the text directly on V .
5
the functional V , which do not necessarily carry over to a general U . Finally,
the model is a special case of the recursive non-expected utility model of
Segal [38]. This will facilitate the comparison of our results with other models
(see, for example, Section 6).
We identify simple lotteries with their cumulative distribution functions,
denoted by capital letters (F, G, and H). Denote by F the set of all cumulative distribution functions of simple lotteries over [x, x]. We assume that V
satisfies the assumptions below (specific conditions on c will be discussed only
in the relevant section). These assumptions are common in the literature on
decision making under risk.
Definition 1 The function V is quasi concave if for any F, G ∈ F and
λ ∈ [0, 1],
V (F ) > V (G) =⇒ V (λF + (1 − λ) G) > V (G).
Quasi concavity implies preference for randomization among equally valued prospects. Together with risk aversion (V (F ) > V (G) whenever G is a
mean preserving spread of F ), quasi concavity implies preference for portfolio diversification (Dekel [15]), which is an important feature when modeling
markets of risky assets.3
Following Machina [28], we assume that V is smooth, in the sense that it
is Fréchet differentiable, defined as follows.
Definition 2 The function V : F → ℜ is Fréchet differentiable if for every
F ∈ F there exists a local utility function uF : [x, x] → ℜ, such that for every
G ∈ F,
Z
V (G) − V (F ) = uF (x)d[G(x) − F (x)] + o(k G − F k)
where k · k is the L1 -norm.
For fixed x > y > z, let (p, q) represent the distribution of the lottery
(z, p; y, 1 − p − q; x, q), where (p, q) ∈ ℜ2+ and p + q 6 1. Such lotteries are
3
The evidence regarding the validity of quasi concavity is supportive yet inconclusive:
while the experimental literature that documents violations of linear indifference curves
(see, for example, Coombs and Huang [14]) found deviations in both directions, that is,
either preference for or aversion to randomization, both Sopher and Narramore [40] and
Dwenger, Kubler, and Weizsacker [17] found explicit evidence in support of quasi concavity.
6
represented in a Marschak-Machina triangle (see panel (i) of Fig. 1). The
bold curves are the indifference curves of V . Quasi concavity implies that
these curves are convex. The dotted parallels to the tangent line at the
distribution F to the indifference curve through this point are indifference
curves of uF (·).
q
q
F
p
p
(i)
(ii)
Figure 1: Indifference curves
The following assumption is taken from Machina [28]:
Definition 3 The Fréchet differentiable functional V satisfies Hypothesis II
if for every x ∈ [x, x],
−
u′′F (x)
u′′G (x)
>
−
u′G (x)
u′F (x)
whenever G first-order stochastically dominates F .
In the Marschak-Machina triangle, Hypothesis II means that indifferu′′ (·)
ence curves are “fanning out” (see panel (ii) of Fig. 1). The term − uF′ (·)
F
is analogues to the Arrow-Pratt measure of risk aversion and has the same
interpretation as in expected utility theory. Hypothesis II was suggested
by Machina as a behavioral regularity that can address known violations of
expected utility, such as the Allais paradox.4
4
While fanning out over some range is necessary to address Allais paradox, the evidence
on global fanning out is mixed. Although fanning out in the lower region of the triangle
is frequently observed, such pattern is inconsistent with many studies that document
“fanning in” indifference curves in the other region. See Camerer [8] for a comprehensive
survey of these findings.
7
For the purpose of our analysis, we only need a weaker notion of Hypothesis II, which requires that property to hold just for degenerate lotteries (i.e.,
Dirac measures), denoted by δy . Formally,
Definition 4 The Fréchet differentiable functional V satisfies Weak Hypothesis II if for every x and for every y > z,
−
2.2
u′′δy (x)
u′δy (x)
>−
u′′δz (x)
.
u′δz (x)
Skewed distributions
Our aim in this paper is to analyze attitude to skewed noise, that is, to
noise that is not symmetric around its mean. For that we need first to
formally define the notion of a skewed distribution. To simplify notation
and terminology we only analyze skewness to the left, but all definitions and
results can be made with skewness to the right.
For a distribution F on [x, x] with expected value µ and for δ > 0, let
R µ−δ
• η1 (F, δ) = x F (x)dx.
Rx
• η2 (F, δ) = µ+δ [1 − F (x)]dx.
Definition 5 The lottery X with the distribution F on [x, x] and expected
value µ is skewed to the left (or negatively skewed) if for every δ > 0,
η1 (F, δ) > η2 (F, δ), that is, if the area below F between x and µ − δ is larger
than the area above F between µ + δ and x (see Fig. 2).
The usefulness of this new notion of skewness will become clear in Section
3.1, where we discuss the proof of our main behavioral result. For now we
only demonstrate that it is stronger than a possible alternative definition,
according to which the lottery
the distribution F and expected value
R x X with
3
µ is skewed to the left if x (y − µ) dF (y) 6 0.
Proposition 1 If X with distribution F and Rexpected value µ is skewed to
x
the left as in Definition 5, then for all odd n, x (y − µ)n dF (y) 6 0.5
5
The converse of Proposition 1 is false. For example, let F be the distribution of
1
1
4
2
the
lottery
(−10,
10 ; −2, 2 ; 0, 35; 7, 7 ). Note that its expected value µ is zero. Moreover,
3
E (X − µ) = −6 < 0 and E (X − µ)2n+1 is decreasing with n, which means that all
odd moments of F are negative. Nevrtheless, the area below the distribution from −10 to
−5 is 12 , but the area above the distribution from 5 to 10 is 47 > 12 , which means that F
is not skewed to the left according to Definition 5.
8
η2 (F, δ)
η1 (F, δ)
x
µ
µ−δ
µ+δ
x
Figure 2: Definition 5, η1 (F, δ) > η2 (F, δ)
3
Asymmetric noise
We now discuss the main topic of the paper, namely the attitude of decision
makers to skewed noise. Recall our notation for two-stage lotteries of the
form hp1 , q1 ; . . . ; pm , qm i, where pi = (x, pi ; x, 1 − pi ) and x > x.
Definition 6 We say that the decision maker rejects symmetric noise if for
all p ∈ (0, 1), for all α 6 min{p, 1 − p}, and for all ε 6 12 ,
hp, 1i hp − α, ε; p, 1 − 2ε; p + α, εi.
As before, we assume that the preference relation over L2 can be represented as in eq. (1) by U (hp1 , q1 ; ...; pn , qn i) = V ((cp1 , q1 ; ...; cpn , qn )), where
V is a functional over simple lotteries and c is a certainty equivalent function.
Theorem 1 Suppose that V is quasi concave, Fréchet differentiable, and
satisfies Weak Hypothesis II. If the decision maker
P always rejects symmetric
noise, then hp, 1i hp1 , q1 ; ...; pn , qn i whenever i pi qi = p and the distribution of (p1 , q1 ; ...; pn , qn ) is skewed to the left.
Furthermore, there are V and c such that the decision maker rejects all
symmetric and negatively skewed noise, yet accepts some positively skewed
noise (i.e., prefers it to hp, 1i).
9
The first part of Theorem 1 provides conditions under which if the decision
maker always rejects symmetric noise, then he will also reject negativelyskewed noise . The conditions on V are familiar in the literature and, as we
have pointed out in the introduction and will further discuss in Section 5,
rejection of symmetric noise is empirically supported. The theoretical link
between attitudes toward symmetric and negatively skewed noise will be
useful in the applications we consider in subsequent sections. The second part
suggests that such behavior is consistent with preference for some positivelyskewed noise. The distinction between positive and negative skewness is
the basis for our analysis, and as we argue, is also supported by empirical
evidence. It is this part of the theorem that distinguishes our model from
other known preferences over compound lotteries that cannot accommodate
rejections of all symmetric noise with acceptance of some positively-skewed
noise (Section 6).
In Example 1 below we go further and introduce a family of functionals for
which we can provide sufficient conditions for acceptance of some positively
skewed noise. In particular, for every p > 0, if the probability q of receiving
(x, p; x, 1 − p) is sufficiently small, then the decision maker will prefer the
noise hp, q; 0, 1 − qi over receiving the lottery (x, pq; x, 1 − pq) for sure. To
guarantee this property, we show that for the functional form of this example,
the first non-zero derivative of V (cpq ), 1) − V (cp , q; 0, 1 − q) with respect to
q at q = 0 is negative (see Appendix A). Note that while Theorem 1 is
independent of the function c, the specification of c is crucial for this result.
ζ
Example 1 Let V (cp1 , q1 ; ...; cpn , qn ) = E[w(cp )]×E[cp ], where w(x) = ζx−x
ζ−1
and cp = βp + (1 − β)pκ .6 These functions satisfy all the assumptions of
Theorem 1, and there is an open neighborhood of (β, ζ, κ) ∈ ℜ3 for which for
every p > 0 there exists a sufficiently small q > 0 such that hp, q; 0, 1 − qi hpq, 1i.
Theorem 1 does not restrict the location of the skewed distribution, but
it is reasonable to find skewed to the left distributions over the value of the
probability p when p is high, and skewed to the right distributions when
p is low. The theorem is thus consistent with the empirical observation
that decision makers reject skewed to the left distributions concerning high
probability of a good event, but seek such distributions when the probability
of the good event is low.
6
The function V belongs to the quadratic utility model of Chew, Epstein, and Segal [10].
10
Our results can explain some of the findings in a recent paper by Abdellaoui, Klibanoff, and Placido [1]. For three different compound lotteries, subjects were asked for their compound lottery premium (as in Dillenberger [16]),
that is, the maximal amount they are willing to pay to replace a compound
lottery with its binary, single-stage counterpart. The underlying binary lottery yields e 50 with probability p and 0 otherwise. The three two-stage
5
5
lotteries were h0.5, 61 ; 0, 56 i, h1, 22
; 0.5, 12
; 0, 22
i, and h1, 65 ; 0.5, 16 i, with base
22
1
, p = 12 , and p = 11
, respectively. They found
probabilities of winning p = 12
12
that the compound lottery premium is an increasing function of p. Other
studies too provide evidence for the pattern of more compound risk aversion
for high probabilities than for low probabilities, and even for compound risk
seeking for low probabilities (see, for example, Kahn and Sarin [24] and Viscusi and Chesson [42]). Following Theorem 1, we argue that it is not only
the magnitude of the probabilities that drive their results, but the fact that
in the three lotteries above, noise is positively skewed, symmetric, and negatively skewed, respectively. Indeed, in a recent paper, Masatlioglu, Orhun,
and Raymond [30] found that individuals exhibit a strong preference for positively skewed noise over negatively skewed ones.
3.1
Outline of the proof of Theorem 1
We discuss only the first part of the theorem, according to which rejection
of symmetric noise implies rejection of negatively skewed noise (the second
part is proved by showing that the functional form in Example 1 satisfies
all the required properties). In the recursive model, rejection of symmetric
noise implies that for any p, the local utility of V at δcp prefers hp, 1i to
hp − a, 12 ; p + a, 12 i. By Weak Hypothesis II, this ranking prevails also when
evaluated using the local utility at δcp∗ , for p∗ > p. Pick a lottery hp, 1i. By
Theorem 2 below, any skewed to the left noise Q around p can be obtained as
the limit of left symmetric splits (Definition 7). By Weak Hypothesis II, each
such split will be rejected when evaluated using the local utility at δcp and,
by Fréchet differentiability, Q itself will also be rejected. Quasi concavity
then implies that it will be rejected globally, that is, hp, 1i Q.
We conclude by stating Theorem 2, which is a mathematical result of an
independent interest.
Definition 7 Let µ be the expected value of a lottery X. Lottery Y is obtained from X by a left symmetric split if Y is the same as X, except for
11
that one of the outcomes x of X such that x 6 µ was split into x + α and
x − α, each with half of the probability of x.
Theorem 2
1. If the lottery Y = (y1 , p1 ; . . . ; yn , pn ) with expected value µ is skewed
to the left, then there is a sequence of lotteries Xi , each with expected
value µ, such that X1 = (µ, 1), Xi → Y , and Xi+1 is obtained from Xi
by a left symmetric split. Moreover, it can be done such that in each
step the size of the spread is bounded by maxi yi − µ.
2. Consider the sequence {Xi } of lotteries where X1 = (µ, 1) and Xi+1 is
obtained from Xi by a left symmetric split. Then the distributions Fi
of Xi converge and the limit distribution F is skewed to the left.
To illustrate the theorem, consider two examples, one where the procedure
terminates in a finite number of steps and one where it does not. In both
cases we move from a degenerate lottery X to a skewed to the left binary
lottery Y with the same expected value as X. For the first example, let
X = (3, 1) and Y = (0, 14 ; 4, 34 ) and obtain
X = (3, 1) → (2, 12 ; 4, 12 ) → (0, 14 ; 4, 14 + 21 ) = Y .
For a sequence that does not terminate, let X = (5, 1) and Y = (0, 61 ; 6, 56 ).
Here we obtain
X = (5, 1) → (4, 12 ; 6, 12 ) → (2, 41 ; 6, 34 ) → (0, 18 ; 4, 18 ; 6, 34 ) → . . .
P
P
(0, 12 n1 41i ; 4, 2·41 n ; 6, 12 + n1 41i ) → . . . (0, 16 ; 6, 56 ) = Y .
The main difficulty in proving part 1 of Theorem 2 is the fact that whereas
outcomes to the left of µ can be manipulated, any split of probabilities that
lands some probability to the right of µ must hit its exact place according to
Y , as we will not be able to touch it later again.
Remark 1 The two parts of Theorem 2 do not create a simple if and only
if statement, because the support of the limit distribution F in part 2 need
not be finite. On the other hand, part 1 of the theorem does not hold for
continuous distributions. By the definition of left symmetric splits, if the
probability of x > µ in Xi is p, then for all j > i, the probability of x
in Xj must be at least p. It thus follows that the distribution F cannot
12
be continuous above µ. However, it can be shown that if F with expected
value µ is skewed to the left, then there is a sequence of finite skewed to the
left distributions Fn , each with expected value µ, such that Fn → F . This
enables us to use Theorem 2 even for continuous distributions.
Remark 2 Menezes, Geiss, and Tressler [31] characterize a notion of increasing downside risk by combining a mean-preserving spread of an outcome
below the mean followed by a mean-preserving contraction of an outcome
above the mean, in a way that the overall result is a transfer of risk from the
right to the left of a distribution, keeping the variance intact. Distribution
F has more downsize risk than distribution G if one can move from G to F
in a sequence of such mean-variance-preserving transformations. Menezes et
al. [31] do not provide a definition (and a characterization as in our Theorem 2) of a skewed to the left distribution. Observe that our characterization
involves a sequence of only symmetric left splits, starting in the degenerate
lottery that puts all the mass on the mean. In particular, our splits are not
mean-variance-preserving and occur only in one side of the mean.
4
Allocation Mechanisms
In this section we apply our results to the comparison of two known allocation mechanisms of indivisible goods. We demonstrate that agents with
preferences as studied in this paper may systematically prefer one mechanism to the other, even though both mechanisms are considered to be the
same in standard models, in the sense that they induce the same probability
distribution over successful matchings.
Consider the following variant of the house allocation problem (Hylland
and Zeckhauser [23]). Let N = {1, 2, ..., n} be a group of individuals and
assume that there are n goods to be allocated among them. The goods are
of two types, g1 and g2 , and we denote by t1 and t2 = n − t1 the number
of units of each type. Each of the n individuals has the same stochastic
preferences, where with probability q he prefers g1 to g2 (independently of
the preferences of other group members). We normalize payoffs so that the
utility from the desired outcome is 1 and the utility from the other outcome
is zero.
Many important goods are allocated using randomizing devices. These
include, among others, the allocation of public schools, course schedule, or
13
dormitory rooms to students, and shifts, offices, or tasks to workers. We
consider two familiar mechanisms, each consists of two stages.
• Random Top Cycle (TC): In the first stage, the allocation of the goods
among the agents is randomly determined, so that for j = 1, 2, the
t
probability of person i to hold good of type gj is nj . In the second stage,
the entire profile of preferences is revealed. Those who like their holding
will keep it. The rest will trade according to the following schedule: If
k people holding one type of good and ℓ 6 k people holding the other
type are unhappy with their holdings, then the latter ℓ will trade and
get their desired outcome, while ℓ out of the former k will be selected
at random and get their preferred option. The other k − ℓ will keep
their undesired outcome.7
• Random Serial Dictatorship (SD): In the first stage the order of the
agents is randomly determined, so that the probability of person i to
be in place j = 1, . . . , n is n1 . In the second stage, the entire profile of
preferences is revealed. The agents then choose the goods according to
the order determined in the first stage. A person will get his desired
outcome if when his turn arrives such a unit is still available.
We adopt an ex ante perspective. In the TC procedure, person i is facing
t
the following two stage lottery. With probability nj he will receive good j.
Once he knows his holding, he will be able to compute the probability that
he’ll eventually obtain his desired outcome, denoted rj . Using our notation,
he’ll face the lottery over probabilities hr1 , tn1 ; r2 , tn2 i. Similarly, once a person
knows his position j in the queue in the SD procedure, he’ll be able to
calculate the probability he’ll win his desired outcome sj . This procedure
thus translates into the lottery hs1 , n1 ; . . . ; sn , n1 i.
The literature on one-sided and two-sided matching (for a recent survey,
see Abdulkadiroğlu and Sönmez [3]) typically maintains the assumption that
agents are only interested in the overall probability they’ll receive their desired outcome. This leads to some results, showing the equivalence of different randomized mechanisms (Abdulkadiroğlu and Sönmez [2]; see also Pathak
and Sethuraman [33]). In particular, Abdulkadiroğlu and Sönmez’s [2] results
7
This is a variant of the classic top cycle mechanism. It can, equivalently, be formulated
more closely to the familiar top cycle, as a problem of matching with indifferences and
using a specific tie-breaking rule. Since the environment we consider is simple, we maintain
our formulation and slightly abuse the title “cycle.”
14
imply that both TC and
P SD lead to the same overall probability of success,
that is, r1 t1 + r2 t2 = j sj .
Recently, Budish et al. [7] pointed out that even if two lotteries are exante equivalent, one may still prefer one to the other because of the nature
of the probabilities they assign to different ex-post outcomes; for example,
one may wish to control for envy ex-post. We show here that our analysis
of preferences for lotteries over probabilities imply some clear cut results
regarding preferences over the aforementioned mechanisms.
Suppose for example that q = 12 and g1 = g2 . This of course does not
mean that everyone will be satisfied, which will be the case only if exactly
half of the population prefer each option. Under TC, receiving g1 or g2 yields
exactly the same lottery and hence the same probability of getting the desired
outcome. TC therefore leads to the degenerate lottery over probabilities
hp, 1i, for some p ∈ (0, 1). Under RD, the first half of the group will be able
to choose their desired outcomes regardless of their preferences. The next
person cannot be certain that his desires will be satisfied, as it may happen
that he and everyone above him will have the same preferences. Denote the
obtained distribution over probabilities by F . As before (see [2]), its expected
value is p. Since for every δ, 1 − F (p + δ) > 12 > F (p − δ), it follows that F
is skewed to the left. By Theorem 1, TC is preferred to RD. Observe that
for this result we didn’t need to compute the exact lotteries induced by TC
and SD.
In the formal analysis to follow, we confine attention to the case of large
(continuum) economies. Let p be the proportion of the g1 units and assume,
without loss of generality, that p > 12 . As before, let q be the probability
that an individual prefers g1 to g2 .
Consider first the case p < q. In the TC mechanism, each individual
receives g1 with probability p, hence qp individuals will receive g1 and will
be satisfied with it. The other (1 − q)p who received g1 will try to replace it
with g2 . Of the 1 − p who received g2 , q(1 − p) will try to replace it with g1 .
Since q > p, (1 − q)p < q(1 − p) and therefore the (1 − q)p individuals who
want to replace g1 with g2 will be able to do so. In other words, all those
who received g1 are satisfied. Of the 1 − p who received g2 , (1 − q)(1 − p) will
keep it, and out of the other q(1 − p), (1−q)p
will be able to replace g2 with
q(1−p)
g1 . Their probability of success is therefore
(1 − q) + q
(1 − q)p
(1 − q)
=
< 1.
q(1 − p)
(1 − p)
15
The TC mechanism thus yields the lottery over probabilities of receiving the
(1−q)
desired outcome given by X1 = h1, p; (1−p)
, 1 − pi.
Consider now the SD mechanism (still assuming p < q). Out of the first
p
, q × pq = p will choose g1 and (1 − q) × pq < 1 − p will choose g2 . As g1
q
will be able to satisfy their desires
is exhausted by the first pq , the other q−p
q
only if they prefer g2 to g1 . The probability of this event is 1 − q. SD thus
leads to the lottery over probabilities given by Y1 = h1, pq ; 1 − q, q−p
i.
q
Similarly, if p > q then TC leads to the lottery X2 = h1, 1 − p; pq , pi and
; q, p−q
i. Observe that independently of the relative
SD leads to Y2 = h1, 1−p
1−q
1−q
magnitude of p and q, the expected probability of receiving the preferred
good is the same under both mechanisms. Also observe that if p = q, then
in a large economy both TC and SD will yield (almost) everyone his desired
outcome for sure.
For the main result in this section, we look at preferences that satisfy the
assumptions of Theorem 1 and in addition satisfy the following property:
Definition 8 The relation displays aversion to larger left-symmetric noise
if for every pi below the average probability p̄, which is obtained with probability qi , and for every ε′ > ε > 0,
. . . ; pi − ε, q2i ; pi + ε, q2i ; . . . . . . ; pi − ε′ , q2i ; pi + ε′ , q2i ; . . . .
We assumed that p > 12 . It is easy to verify that for all q > p, Y1 is a mean
preserving spread of X1 , and for all p ≥ q, Y2 is a mean preserving spread
of X2 . As argued before, this may not be enough to guarantee preference
for the X lotteries over the Y lotteries. The next theorem analyzes these
preferences.
Theorem 3 Suppose that satisfies the assumptions of Theorem 1 and displays aversion to larger left-symmetric noise (Definition 8). If q > 2p − 1,
then in large economies TC is preferred to SD.
(1−q)
Since p > 12 , X1 = h1, p; (1−p)
, 1 − pi is skewed to the left and X2 =
q
h1, 1 − p; p , pi is skewed to the right. For p < q < 1, pq > 12 and therefore
i is skewed to the left. As long as 1−p
> 21 , Y2 =
Y1 = h1, pq ; 1 − q, q−p
q
1−q
; q, p−q
i is skewed to the left. However, when p > 2p − 1 > q, we
h1, 1−p
1−q
1−q
1−p
obtain that 1−q < 21 and Y2 too is skewed to the right. Consider now the
16
case where p is large and q is small.8 In this case most of the outcomes are
of one type, but individuals are not likely to like it. Both X2 and Y2 are
then lotteries that yield with high probability a bad outcome and a good
outcome with a small probability. Following Theorem 1, we know that there
are functionals that satisfy our assumptions but that for sufficiently small q
and sufficiently large p, both mechanisms will be considered better than a
(hypothetical) mechanism that will offer the desired outcome with the (same)
expected probability 1 − p + q.
The lottery Y2 is a mean preserving spread of X2 . It yields the good
outcome (probability 1) with a higher probability than X2 , but its bad outcome is worse than the bad outcome under X2 . Moving from X2 to Y2 thus
increases the small probability of the good probability (1) at the cost of
(slightly) reducing the desirability of the bad probability. This has the flavor
of our arguments for preference for a lottery that yields with a small probability a good chance of getting a good outcome over a simple lottery that
yields the good with the expected probability. The following proposition
establishes this connection.
Proposition 2 The following two statements are equivalent:
1. For a sufficiently small q, h1, q; 0, 1 − qi hq, 1i.
2. For a sufficiently small q and sufficiently large p, Y2 X2 (that is, SD
is preferred to TC).
Once again, the preferences of Example 1 are consistent with our analysis
(see Appendix A). These preferences satisfy all the assumptions of Theo2p
rem 3, except for the regions 2p − 1 < q < p2 and 1+p
< q < 2p − p2 .
Nevertheless, we show by means of numerical analysis that on these regions
too TC is preferred to SD. Following Proposition 2, for a sufficiently high p
and sufficiently small q, the preferences of Example 1 rank SD above TC.
5
Ambiguity aversion and seeking
Ambiguity aversion is one of the most investigated phenomena in decision
theory. Consider the classic Ellsberg [18] thought experiment: subjects are
8
This case is not covered by Theorem 3, where Case 2.3 in its proof strongly depends
on the assumption that q > 2p − 1.
17
presented with two urns. Urn 1 contains 100 red and black balls, but the exact
color composition is unknown. Urn 2 has exactly 50 red and 50 black balls in
it. Subjects are asked to choose an urn from which a ball will be drawn, and
to bet on the color of this ball. If a bet on a specific urn is correct the subject
wins $100, zero otherwise. Let Ci be the bet on a color (Red or Black) draw
from Urn i. Ellsberg predicted that most subjects will be indifferent between
R1 and B1 as well as between R2 and B2 , but will strictly prefer R2 to R1
and B2 to B1 . While, based on symmetry arguments, it seems plausible that
the number of red balls in urn 1 equals the number of black balls, Urn 1 is
ambiguous in the sense that the exact distribution is unknown whereas urn 2
is risky, as the probabilities are known. An ambiguity averse decision maker
will prefer to bet on the risky urn to bet on the ambiguous one. Ellsberg’s
predictions were confirmed in many experiments.9
The recursive model we study here was first suggested by Segal [37] as a
formal way to capture ambiguity aversion.10 Under this interpretation, ambiguity is identified as two-stage lotteries. The first stage captures the decision
maker’s uncertainty about the true probability distribution over the states
of the world (the true composition of the urn in Ellsberg’s example), and the
second stage determines the probability of each outcome, conditional on the
probability distribution that has been realized. Holding the prior probability
distribution over states fixed, an ambiguity averse decision maker prefers the
objective (unambiguous) simple lottery to any (ambiguous) compound one.
Note that according to Segal’s model, preferences over ambiguous prospects
are induced from preferences over the compound lotteries that reflect the decision maker’s beliefs. That is, the first stage is imaginary and corresponds
to the decision maker’s subjective beliefs over the values of the true probabilities.
While Ellsberg-type behavior seems intuitive and is widely documented,
there are situations where decision makers actually prefer not to know the
probabilities with much preciseness. Becker and Bronwson [4, fnt. 4] describe
9
A comprehensive reference to the evidence on Ellsberg-type behavior, and on attitude
towards ambiguity in general, can be found in Peter Wakker’s annotated bibliography,
posted at his website under http://people.few.eur.nl/wakker/refs/webrfrncs.doc.
10
There are many other ways to model ambiguity aversion. Prominent examples include
Choquet expected utility (Schmeidler [36]), maximin expected utility (Gilboa and Schmeidler [20]), variational preferences (Maccheroni, Marinacci, and Rustichini [29]), α-maxmin
(Ghirardato, Maccheroni, and Marinacci [19])), and the smooth model of ambiguity aversion (Klibanoff, Marinacci, and Mukerji [27]).
18
a conversation where Ellsberg himself suggested that people may prefer ambiguity with respect to a low-probability good event. There is indeed a growing
experimental literature that challenges the assumption that people are globally ambiguity averse. See, among others, Chew, Miao, and Zhong [11], and
van de Kuilen and Wakker [41]. Consider the following situation. A person
suspects that there is a high probability that he will face a bad outcome (severe loss of money, serious illness, criminal conviction, etc.). Yet he believes
that there is some (small) chance things are not as bad as they seem (Federal
regulations will prevent the bank from taking possession of his home, it is
really nothing, they won’t be able to prove it). These beliefs might emerge,
for example, from consulting with a number of experts (such as accountants,
doctors, lawyers) who disagree in their opinions; the vast majority of which
are negative but some believe the risk is much less likely. Does the decision
maker really want to know the exact probabilities of these events? The main
distinction between the sort of ambiguity in Ellsberg’s experiment and the
ambiguity in the last examples is that the latter is asymmetric and, in particular, positively skewed. On the other hand, if the decision maker expects
a good outcome with high probability, he would probably prefer to know this
probability for sure, rather than knowing that there is actually a small chance
that things are not that good. In other words, asymmetric but negatively
skewed ambiguity may well be undesired.
To illustrate attitudes towards skewed ambiguity, reconsider Boiney’s [6]
experiment given in the introduction.11 A decision maker can choose one of
three investment plans, for each of which he believes that the overall probability that the investment be successful is the same, p. The decision maker
prefers an investment plan A, in which he is confident about its probability
of success over plan B that describes a negatively skewed distribution around
p, where it is very likely that the true probability slightly exceeds p but it is
also possible (albeit unlikely) that the true probability is very low. Yet, he
prefers Option C, which describes a positively skewed distribution around
p, where it is very likely that the true probability falls slightly below p but
there is a small chance it is actually a decent probability, over both A and
B.
Our model is consistent with this type of behavior. Let (x, p; x, 1 − p)
11
This example is slightly different from the intuition given in the previous paragraph;
it demonstrates different ambiguity attitude towards different patterns of noise around a
given average value, rather than towards skewed to the right (left) noise around low (high)
average probability.
19
denote a conceivable probability distribution over the two possible states
(success, prize x, and failure, prize x < x). For some q1 , q2 > 21 and p1 <
p2 < p < p3 < p4 such that q1 p3 + (1 − q1 )p1 = q2 p2 + (1 − q2 )p4 = p, the three
investment options A, B, and C can be described as the two-stage lotteries
hp, 1i, hp3 , q1 ; p1 ; 1−q1 i, and hp2 , q2 ; p4 , 1−q2 i, respectively. Since B is skewed
to the left, the Hypothesis of Theorem 1 implies that A is preferred to B.
And since C is skewed to the right, we have no implication for the ranking of
C and A. In particular, the model permits having strict preference for C (for
instance, using the functional form of Example 1), in line with our intuition
and with the observed data.12
6
Relations to the literature
The proofs of the theorems presented in this paper are quite complicated,
and it is therefore natural to ask whether simpler models can deliver the
same results.
Consider first a recursive model in which the decision maker is an expected
utility maximizer in each of the two stages. Let u and v be the vNM utility
functions over outcomes in the first and in the second stages. The decision
maker evaluates a two-stage lottery hp1 , q1 ; ...; pn , qn i recursively by
X
U (hp1 , q1 ; ...; pn , qn i) =
qi u v −1 (Ev [pi ])
(2)
i
−1
where v (Ev [p]) is the certainty equivalent of lottery p calculated using the
function v. In the context of temporal lotteries, this model is a special case
of the one studied by Kreps and Porteus [26]. In the context of ambiguity,
this is the model of Klibanoff, Marinacci, Mukerji [27]. We now argue that
while this model is consistent with the rejection of all symmetric noise, it
cannot accommodate preferences that reject all symmetric noise while still
accepting some positively skewed noise. To see this, note that by eq. (2), for
any p, the value of the noise hp + α, 12 ; p − α, 21 i is
−1
1
u
v
[(p
+
α)v(x)
+
(1
−
p
−
α)v(y)]
+
2
12
Boiney [6] used a two-stage lottery as an operational definition of ambiguity. His
experiment involves prospects like the ones describe in the text, with the specification
x = $200, x = $0, p = 0.2, q1 = q2 = 0.9, p1 = 0.02, p2 = 0.18, p3 = 0.22, and p4 = 0.38.
His data strongly supports the ranking C ≻ A ≻ B. He further showed that this ranking
is robust for a wide range of parameters.
20
1
u
2
v −1 [(p − α)v(x) + (1 − p + α)v(y)] .
The value of the simple lottery (x, p; x, 1 − p) is13
u v −1 [pv(x) + (1 − p)v(x)] .
Rejection of symmetric noise implies that for any p and α in the relevant
range,
u v −1 [pv(x) + (1 − p)v(x)] >
(3)
−1
1
u v [(p + α)v(x) + (1 − p − α)v(x)] +
2
−1
1
u
v
[(p
−
α)v(x)
+
(1
−
p
+
α)v(x)]
.
2
Pick any two numbers a > b in [0, 1] and note that by setting p = 21 (a + b)
and α = 21 (a − b), inequality (3) is equivalent to the requirement that
u v −1 [ 12 (a + b)v(x) + (1 − 21 (a + b))v(x)] >
1
−1
−1
1
+
(1
−
a)v(x)]
+
+
(1
−
b)v(x)]
.
u
v
[av(x)
u
v
[bv(x)
2
2
Since a and b are arbitrary, this inequality should hold for all such pairs. This
is the case if and only if the function u ◦ v −1 is mid-point concave, which by
continuity implies that u◦v −1 is concave. But then the decision maker would
reject any noise.
This feature of the recursive expected utility model implies that it cannot accommodate the pattern of preferences over allocation mechanisms of
Section 4. In particular, since the lottery over probabilities induced by SD is
a mean preserving spread of the one induced by TC, the recursive expected
utility model predicts that independently of the values of p and q, a decision
maker who rejects symmetric noise will always prefer TC to SD. In contrast,
as we establish in Theorem 3 and Proposition 2, while rejection of symmetric noise in our model implies preference for TC over SD for many possible
13
Alternatively, the simple lottery can be translated into the two-stage lottery that pays
δx with probability p and δx with probability 1-p, where δx is the distribution of the lottery
that yields x ∈ [x, x] with probability one. Its recursive value is
pu v −1 [v(x)] + (1 − p)u v −1 [v(x)] = pu(x) + (1 − p)u(x).
The result in this section is independent of this specification; in the analysis below, rejecting symmetric noises would imply that u ◦ v −1 is convex, which again implies rejection of
any noise.
21
combinations of p and q, it does not preclude preferring SD to TC, especially
when the induced lotteries over probabilities are skewed to the right.
Dillenberger [16] studied a property called preferences for one-shot resolution of uncertainty, which, in the language of this paper, means that the
decision maker rejects all noise. Dillenberger confined his attention to recursive preferences over two-stage lotteries in which the certainty equivalent
functions in the second stage are calculated using the same V that applied in
the first stage.14 He showed an equivalence between preferences for one-shot
resolution of uncertainty and a static property called negative certainty independence (NCI). Recently, Cerreia Vioglio, Dillenberger, and Ortoleva [9]
derived a utility representation — termed cautious expected utility — of all
preferences that satisfy NCI together with basic rationality postulates.
Our analysis in this paper is based on functionals V that are Fréchet differentiable. Segal and Spivak [39] defined a preference relation as exhibiting
second-order (resp., first-order) risk aversion if the derivative of the implied
risk premium on a small, actuarially fair gamble vanishes (resp., does not
vanish) as the size of the gamble converges to zero.15 If V is Fréchet differentiable, then it satisfies second-order risk aversion. There are interesting
preference relations that are not Fréchet differentiable (and that do not satisfy second-order risk aversion), e.g., the rank-dependent utility model of
Quiggin [34]. The analysis of attitudes towards skewed noise in the recursive model for general first-order risk aversion preferences is not vacuous,
but is significantly different than the one presented in this paper and will be
developed in future work.
Appendix: Proofs
Proof of Proposition 1: Let the lottery Y be obtained from the lottery Z
by a left symmetric split and denote by x their common mean. For example,
the outcome zi 6 x with probability pi of Z is split into zi −α and zi +α, each
with probability p2i . Denote the distributions of Y and Z by F and G. Since
for t < 0 and odd n, tn is a concave function, it follows that if zi + α 6 x,
14
This is known as the time neutrality axiom (Segal [38]).
Formally, if π(t) is the amount of money that an agent would pay to avoid the nondegenerate gamble x + te
ε, where E (e
ε) = 0, then π(t) is O(t) and o(t) for first and secondorder risk averse preferences, respectively.
15
22
then
Z
x
n
(t − x) dF (t) −
x
Z
x
(t − x)n dG(t) =
x
pi
[(zi − α − x)n + (zi + α − x)n ] − pi (zi − x)n 6 0.
2
(4)
If zi + α > x we need to manipulate eq. (4) a little further. Let ξ = zi − x
and obtain
pi
[(zi − α − x)n + (zi + α − x)n ] − pi (zi − x)n =
2
pi
[(ξ − α)n + (ξ + α)n ] − pi ξ n =
2
n−1
n−1
2 2 X
p
pi n pi X
n
n
i
2j−1 n−2j+1
ξ
α
−
ξ 2j αn−2j +
ξ +
2j
−1
2j
2
2 j=1
2 j=0
n−1
n−1
2 2 X
pi n pi X
p
n
n
i
2j−1 n−2j+1
ξ +
ξ
α
+
ξ 2j αn−2j − pi ξ n =
2
2 j=1 2j −1
2 j=0 2j
2 X
n
ξ 2j−1 αn−2j+1 6 0
pi
2j −1
n−1
j=1
where the last inequality follows by the fact that ξ 6 0. Since X with
expected value µ is skewed to the left it follows by Theorem 2 that it can
be obtain as the limit of a sequence of left symmetric splits. At δµ (the
Rx
distribution of (µ, 1)), x (y − µ)n dδµ = 0. The claim follows by the fact that
each left symmetric split reduces the value of the integral.
Proof of Theorem 1: Let cp be the certainty equivalent of the lottery
(x, p; x, 1 − p). The two-stage lottery hp − α, ε; p, 1 − 2ε; p + α, εi translates
in the recursive model into the lottery (cp−α , ε; cp , 1 − 2ε; cp+α , ε). Since the
decision maker always rejects symmetric noise, it follows that the local utility
uδcp satisfies
uδcp (cp ) > 12 uδcp (cp−α ) + 12 uδcp (cp+α ).
By Weak Hypothesis II, for every r > p,
uδcr (cp ) > 21 uδcr (cp−α ) + 21 uδcr (cp+α ).
23
(5)
Consider first
P the lottery over the probabilities given by Q = hp1 , q1 ; . . . ;
pm , qm i where qi pi = p (we deal with the distributions with non-finite support at the end of the proof) . If Q is skewed to the left, then it follows by Theorem 2 that there is a sequence of lotteries Qi = hpi,1 , qi,1 , . . . , pi,ni , qi,ni i → Q
such that Q1 = hp, 1i and for all i, Qi+1 is obtained from Qi by a left symmetric split. For each i, let Q̃i = (cpi,1 , qi,1 ; . . . ; cpi,ni , qi,ni ). Suppose pi,j is
split into pi,j − α and pi,j + α. By eq. (5), as p > pi,j ,
E[uδp (Q̃i )] =
qi,j uδcp (cpi,j ) +
X
qi,m uδcp (cpi,m ) >
m6=j
1
)
q u (c
2 i,j δcp pi,j −α
+ 12 qi,j uδcp (cpi,j +α ) +
X
qi,m uδcp (cpi,m ) =
m6=j
E[uδcp (Q̃i+1 )].
As Qi → Q, and as for all i, uδcp (cp ) > E[uδp (Q̃i )], it follows by continuity
that uδcp (cp ) > E[uδcp (Q̃)]. By Fréchet Differentiability
∂ V εQ̃ + (1 − ε)δcp 6 0.
∂ε
ε=0
Quasi-concavity now implies that V (δcp ) > V (Q̃), or hp, 1i Q. Finally, as
preferences are continuous, it follows by Remark 1 that the theorem holds
for all Q, even if its support is not finite.
For the second part of the theorem, we show in Appendix A that the
functional form in Example 1 satisfies all the assumptions of the theorem
and always accepts some positively skewed noise.
Proof of Theorem 2: Lemma 1 proves part 1 of the theorem for binary
lotteries Y . After a preparatory claim (Lemma 2), the general case of this
part is proved in Lemma 3 for lotteries Y with FY (µ) > 12 , and for all
lotteries in Lemma 4. That this can be done with bounded shifts is proved
in Lemma 5. Part 2 of the theorem is proved in Lemma 6.
Lemma 1 Let Y = (x, r; z, 1 − r) with mean E[Y ] = µ, x < z, and r 6
1
. Then there is a sequence of lotteries Xi with expected value µ such that
2
X1 = (µ, 1), Xi → Y , and Xi+1 is obtained from Xi by a left symmetric split.
Moreover, if ri and ri′ are the probabilities of x and z in Xi , then ri ↑ r and
ri′ ↑ 1 − r.
24
Proof: The main idea of the proof is to have at each step at most five
outcomes: x, µ, z, and up to two outcomes between x and µ. In a typical
move either µ or one of the outcomes between x and µ, denote it w, is split
“as far as possible,” which means:
], then split its probability between x and w + (w − x) =
1. If w ∈ (x, x+µ
2
2w − x. Observe that x < 2w − x 6 µ.
2. If w ∈ [ x+z
, µ], then split its probability between z and w − (z − w) =
2
2w − z. Observe that x 6 2w − z < µ.
, x+z
), then split its probability between µ and w−(µ−w) =
3. If w ∈ ( x+µ
2
2
2w − µ. Observe that x < 2w − µ < µ.
If r = 12 , that is, if µ = x+z
then the sequence terminates after the first split.
2
We will therefore assume that r < 21 . Observe that the this procedures never
split the probabilities of x and z hence these probabilities form increasing
sequences. We identify and analyze three cases: a. For every i the support
of Xi is {x, yi , z}. b. There is k > 1 such that the support of Xk is {x, µ, z}.
c. Case b does not happen, but there is k > 1 such that the support of Xk
is {x, wk , µ, z}. We also show that if for all i > 1, µ is not in the support of
Xi , then case a prevails.
a. The simplest case is when for every i the support of Xi has three outcomes
at most, x < yi < z. By construction, the probability of yi is 21i , hence Xi
puts 1 − 21i probability on x and z. In the limit these converge to a lottery
over x and z only, and since for every i, E[Xi ] = µ, this limit must be Y .
The two examples of Section 2.2 show that this procedure may or may not
terminate after a finite number of steps.
b. Suppose now that even though at a certain step the obtained lottery has
more than three outcomes, it is nevertheless the case that after k splits we
reach a lottery of the form Xk = (x, pk ; µ, qk ; z, 1 − pk − qk ). For example,
7
let X = (17, 1) and Y = (24, 17
; 0, 24
). The first five splits are
24
X = (17, 1) → (10, 21 ; 24, 12 ) → (3, 14 ; 17, 14 ; 24, 12 ) →
3
1
(0, 18 ; 6, 18 ; 17, 14 ; 24, 12 ) → (0, 16
; 12, 16
; 17, 41 ; 24, 12 ) →
17
7
; 17, 14 ; 24, 32
)
(0, 32
25
(6)
By construction k > 2 and qk 6 14 . Repeating these k steps j times will yield
the lottery Xjk = (x, pjk ; µ, qkj ; z, 1 − pjk − qkj ) → Y as qkj → 0 and as the
expected value of all lotteries is µ, pjk ↑ r and 1 − pjk − qji ↑ 1 − r.
c. If at each stage Xi puts no probability on µ then we are in case a. The
reason is that as splits of type 3 do not happen, in each stage the probability
of the outcome between x and z is split between a new such outcome and
either x or z, and the number of different outcomes is still no more than
three. Suppose therefore that at each stage Xi puts positive probability on
at least one outcome w strictly between x and µ (although these outcomes
w may change from one lottery Xi to another) and at some stage Xi puts
(again) positive probability on µ. Let k > 2 be the first split that puts
positive probability on µ. We consider two cases.
c1 . k = 2: In the first step, the probability of µ is divided between z and
2µ − z and in the second step the probability of 2µ − z is split and half of
it is shifted back to µ (see for example the second split in eq. (6) above). In
other words, the first split is of type 2 while the second is of type 3. By the
description of the latter,
x+µ
x+z
2
µ−x
3
< 2µ − z <
⇐⇒ <
<
2
2
3
z−x
4
(7)
The other one quarter of the original probability of µ is shifted from 2µ − z
to
x+µ
⇐⇒ 4(z − x) > 5(µ − x)
2µ − z − (µ − [2µ − z]) = 3µ − 2z 6
2
Which is satisfied by eq. (7). Therefore, in the next step a split of type 1 will
be used, and one eighth of the original probability of µ will be shifted away
from 2µ − z to x. In other words, in three steps 85 of the original probability
of µ is shifted to x and z, one quarter of it is back at µ, and one eighth of it
is now on an outcome w1 < µ.
⋄
19
c2 . k > 3: For example, X = (29, 1) and Y = (48, 29
; 0, 48
). Then
48
X = (29, 1) → (10, 12 ; 48, 12 ) → (0, 14 ; 20, 14 ; 48, 12 ) →
(0, 41 ; 11, 18 ; 29, 18 ; 48, 21 ) → . . .
(8)
After k splits 21k of the original probability of µ is shifted back to µ and 21k
is shifted to another outcome w1 < µ. The rest of the original probability is
split (not necessarily equally) between x and z.
⋄
26
Let ℓ = max{k, 3}. We now construct inductively a sequence of cycles,
where the length of cycle j is ℓ + j − 1. Such a cycle will end with the
probability distributed over x < wj < µ < z. Denote the probability of µ by
pj and that of wj by qj . We show that pj + qj → 0. The probabilities of x
and z are such that the expected value is kept at µ, and as pj + qj → 0, it
will follow that the probabilities of x and z go up to r and 1 − r, respectively.
In the example of eq. (8), ℓ = 3, the length of the first cycle (where j = 1) is
3, and w1 = 11.
Suppose that we’ve finished the first j cycles. Cycle j + 1 starts with
splitting the pj probability of µ to {x, w1 , µ, z} as in the first cycle. One
of the outcomes along this sequence may be wj , but we will continue to
split only the “new” probability of this outcome (and will not yet touch
the probability qj of wj ). At the end of this part of the new cycle, the
probability is distributed over x, w1 , wj , µ, and z. At least half of pj , the
earlier probability of µ, is shifted to {x, z}, and the probabilities of both
these outcomes did not decrease. Continuing the example of eq. (8), the first
part of the second cycle (where j = 1) is
1
9
(0, 41 ; 11, 18 ; 29, 18 ; 48, 12 ) → (0, 14 ; 10, 16
; 11, 81 ; 48, 16
)→
9
1
1
9
9
9
1
9
(0, 32 ; 11, 8 ; 20, 32 ; 48, 16 ) → (0, 32 ; 11, 64 ; 29, 64 ; 48, 16
)
The second part of cycle j + 1 begins with j − 1 splits starting with w1 .
At the end of these steps, the probability is spread over x, wj , µ, and z. Split
the probability of wj between an element of {x, µ, z} and wj+1 which is not
in this set to get pj+1 and qj+1 . In the above example, as j = 1 there is only
one split at this stage to
45
9
1
9
(0, 128
; 22, 128
; 29, 64
; 48, 16
)
And w2 = 22. The first part of the third cycle (j = 2) leads to
91
1
9
1
73
; 11, 512
; 22, 128
; 29, 512
; 48, 128
)
(0, 256
The second part of this cycle has two splits. Of w1 = 11 into 0 and 22, and
then of w2 = 22 into µ = 29 and w3 = 15.
73
1
73
365
73
77
73
365
; 22, 1024
; 29, 512
; 48, 128
) → (0, 1024
; 15, 2048
; 29, 2048
; 48, 128
)
→ (0, 1024
We now show that for every j,
pj+2 + qj+2 6 34 (pj + qj )
27
(9)
We first observe that for every j, pj+1 + qj+1 < pj + qj . This is due to the
fact that the rest of the probability is spread over x and z, the probability
of z must increase (because of the initial split in the probability of µ), and
the probabilities of x and z cannot go down.
When moving from (pj , qj ) to (pj+2 , qj+2 ), half of pj is switched to z.
Later on, half of qj is switched either to x or z, or to µ, in which case half of
it (that is, one quarter of qj ) will be switched to z on the move from pj+1 to
pj+2 . This proves inequality (9), hence the lemma.
Lemma 2 Let X = (x1 , p1 ; . . . ; xn , pn ) and Y = (y1 , q1 ; . . . ; ym , qm ) where
x1 6 . . . 6 xn and y1 6 . . . 6 ym be two lotteries such that X dominates Y
by second-order stochastic dominance. Then there is a sequence of lotteries
Xi such that X1 = X, Xi → Y , Xi+1 is obtained from Xi by a symmetric
(not necessarily always left or always right) split of one of the outcomes of
Xi , all the outcomes of Xi are between y1 and ym , and the probabilities the
lotteries Xi put on y1 and ym go up to q1 and qm , respectively.
Proof: From Rothschild and Stiglitz [35,
that we can present
P p. 236] we know P
Y as (y11 , q11 ; . . . ; ynn , qnn ) such that j qkj = pk and j qkj ykj /pk = xk ,
k = 1, . . . , n.
Let Z = (z1 , r1 ; . . . ; zℓ , rℓ ) such that z1 < . . . < zℓ and E[Z] = z. Let
Z0 = (z, 1). One can move from Z0 to Z in at most ℓ steps, where at each
step some of the probability of z is split into two outcomes of Z without
affecting the expected value of the lottery, in the following way. If
r1 z 1 + rℓ z ℓ
>z
r1 + rℓ
(10)
then move r1 probability to z1 and rℓ′ 6 rℓ to zℓ such that r1 z1 + rℓ′ zℓ = z(r1 +
rℓ′ ). However, if the sign of the inequality in (10) is reversed, then move rℓ
probability to zℓ and r1′ 6 r1 probability to zℓ such that r1′ z1 +rℓ zℓ = z(r1′ +rℓ ).
Either way the move shifted all the required probability from z to one of the
outcomes of Z without changing the expected value of the lottery.
Consequently, one can move from X to Y in ℓ2 steps, where at each step
some probability of an outcomes of X is split between two outcomes of Y .
By Lemma 1, each such split can be obtained as the limit of symmetric splits
(recall that we do not require in the current lemma that the symmetric splits
will be left or right splits). That all the outcomes of the obtained lotteries
are between y1 and ym , and that the probabilities these put on y1 and ym go
up to q1 and qm follow by Lemma 1.
28
Lemma 3 Let Y = (y1 , p1 ; . . . ; yn , pn ), y1 6 . . . 6 yn , with expected value
µ be skewed to the left such that FY (µ) > 12 . Then there is a sequence of
lotteries Xi with expected value µ such that X1 = (µ, 1), Xi → Y , and Xi+1
is obtained from Xi by a left symmetric split. Moreover, if ri and ri′ are the
probabilities of y1 and yn in Xi , then ri ↑ p1 and ri′ ↑ pn .
Proof: Suppose wlg that yj ∗ = µP(of course, it may be that pj ∗ = 0). Since
FY (yj ∗ ) > 12 , it follows that t := nj=j ∗+1 pj 6 21 . As Y is skewed to the left,
yn − µ 6 µ − y1 , hence 2µ − yn > y1 . Let m = n − j ∗ be the number of
outcomes of Y that are strictly above the expected value µ. Move from X1 to
Xm = (2µ − yn , pn ; . . . ; 2µ − yj ∗ +1 , pj ∗ +1 ; yj ∗ , 1 − 2t; yj ∗ +1 , pj ∗ +1 ; . . . ; yn , pn ) by
repeatedly splitting probabilities away from µ. All these splits are symmetric,
hence left symmetric splits.
Next we show that Y is a mean preserving spread of Xm . Obviously,
E[Xm ] = E[Y ] = µ. Integrating by parts, we have for x > µ
Z yn
Z x
Z yn
yn − µ =
FY (z)dz =
FY (z)dz
FY (z)dz +
y1
y1
x
Z yn
Z x
Z yn
FXm (z)dz +
FXm (z)dz
FXm (z)dz =
yn − µ =
x
y1
y1
Rx
Since FY and FXm coincide for z > µ, we have, for x > µ, y1 FXm (z)dz =
Rx
Rx
Rx
F (z)dz.
F
(z)dz
6
F
(z)dz
and
in
particular,
X
Y
m
y1 Y
y1
y1
For x < µ it follows by the assumption that Y is skewed to the left and
by the construction of Xm as a symmetric lottery around µ that
Z x
Z 2µ−y1
FXm (z)dz =
[1 − FXm (z)]dz =
y1
Z
2µ−x
2µ−y1
[1 − FY (z)]dz 6
2µ−x
Z
x
FY (z)dz
y1
Since to the right of µ, Xm and Y coincide, we can view the left side of
Y as a mean preserving spread of the left side of Xm . By Lemma 2 the left
side of Y is the limit of symmetric mean preserving spreads of the left side of
Xm . Moreover, all these splits take place between y1 and µ and are therefore
left symmetric splits. By Lemma 2 it also follows that ri ↑ p1 and ri′ ↑ pn . We now show that Lemma 3 holds without the restriction FY (µ) > 12 .
29
Lemma 4 Let Y with expected value µ be skewed to the left. Then there is a
sequence of lotteries Xi with expected value µ such that X1 = (µ, 1), Xi → Y ,
and Xi+1 is obtained from Xi by a left symmetric split.
Proof: The first step in the proof of Lemma 3 was to create a symmetric
distribution around µ such that its upper tail (above µ) agrees with FY .
Obviously this can be done only if FY (µ) > 12 , which is no longer assumed.
Instead, we apply the proof of Lemma 3 successively to mixtures of FY and
δµ , the distribution that yields µ with probability one.
Suppose that FY (µ) = λ < 12 . Let γ = 1/2(1 − λ) and define Z to be the
lottery obtained from the distribution γFY + (1 − γ)δµ . Observe that
FZ (µ) = γFY (µ) + (1 − γ)δµ (µ) =
1 − 2λ
1
λ
+
=
2(1 − λ) 2(1 − λ)
2
It follows that the lotteries Z and (µ, 1) satisfy the conditions of Lemma 3,
and therefore there is sequence of lotteries Xi with expected value µ such
that X1 = (µ, 1), Xi → Z, and Xi+1 is obtained from Xi by a left symmetric
split. This is done in two stages. First we create a symmetric distribution
around µ that agrees with Z above µ (denote the number of splits needed in
this stage by t), and then we manipulate the part of the distribution which is
weakly to the left of µ by taking successive symmetric splits (which are all left
symmetric splits when related to µ) to get nearer and nearer to the secondorder stochastically dominated left side of Z as in Lemma 2. Observe that
the highest outcome of this part of the distribution Z is µ, and its probability
is 1 − γ. By Lemma 2, for every k > 1 there is ℓk such that after ℓk splits of
1
this second phase the probability of µ will be at least rk = (1 − γ)(1 − k+1
)
1
and k Xt+ℓk − Z k< k .
The first cycle will end after t+ℓ1 splits with the distribution FZ1 . Observe
that the probabilities of the outcomes to the right of µ in Z1 are those of
the lottery Y multiplied by γ. The first part of second cycle will be the
same as the first cycle, applied to the rk conditional probability of µ. At the
end of this part we’ll get the lottery Z1′ which is the same as Z1 , conditional
on the probability of µ. We now continue the second cycle by splitting the
combination of Z1 and Z1′ for the total of t + ℓ1 + ℓ2 steps. As we continue to
add such cycles inductively we get closer and closer to Y , hence the lemma. Next we show that part 1 of the theorem can be achieved by using
bounded spreads. The first steps in the proof of Lemma 3 involve shifting probabilities from µ to all the outcomes of Y to the right of µ, and these
30
outcomes are not more than max yi − µ away from µ. All other shifts are
symmetric shifts involving only outcomes to the left of µ. The next lemma
shows that such shifts can be achieved as the limit of symmetric bounded
shifts.
Lemma 5 Let Z = (z−α, 21 ; z+α, 12 ) and let ε > 0. Then there is a sequence
of lotteries Zi such that Z0 = (z, 1), Zi → Z, and Zi+1 is obtained from Zi
by a symmetric (not necessarily left or right) split of size smaller than ε.
Proof: The claim is interesting only when ε < α. Fix n such that ε > αn .
We show that the lemma can be proved by choosing the size of the splits to
be αn . Consider the 2n + 1 points zk = z + nk , k = −n, . . . , n and construct
the sequence {Zi } where Zi = (z − α, pi,−n ; z − n−1
α, pi,−n+1 ; . . . ; z + α, pi,n )
n
as follows.
The index i is odd: Let zj be the highest outcome in {z, . . . , z +
with the highest probability in Zi−1 . Formally, j satisfies:
n−1
α}
n
• 06j 6n−1
• pi−1,j > pi−1,k for all k
• If for some j ′ ∈ {0, . . . , n − 1}, pi−1,j ′ > pi−1,k for all k, then j > j ′ .
Split the probability of zj between zj − αn and zj + αn (i.e., between zj−1 and
zj+1 ). That is, pi,j−1 = pi−1,j−1 + 21 pi−1,j , pi,j+1 = pi−1,j+1 + 21 pi−1,j , pi,j = 0,
and for all k 6= j − 1, j, j + 1, pi,k = pi−1,k .
The index i is even: In this step we create the mirror split of the one done
in the previous step. Formally, If j of the previous stage is zero, do nothing.
Otherwise, split the probability of z−j between z−j − αn and z−j + αn . That
is, pi,−j−1 = pi−1,−j−1 + 12 pi−1,−j , pi,−j+1 = pi−1,−j+1 + 21 pi−1,−j , pi,−j = 0, and
for all k 6= −j − 1, −j, −j + 1, pi,k = pi−1,k .
After each pair of these steps, the probability distribution is symmetric around z. Also, the sequences {pi,−n }i and {pi,n } are non decreasing.
Being bounded by 21 , they converge to a limit L. Our aim is to show
that L = 21 . Suppose not. Then at each step the highest probability of
{pi−1,−n+1 , . . . , pi−1,n−1 must be at least ℓ := (1 − 2L)/(2n − 1) > 0. The
variance of Zi is bounded from above by the variance of (µ − α, 12 ; µ + α, 21 ),
which is α2 . Splitting p probability from z to z − αn and z + αn will increase
31
the variance by p( αn )2 . Likewise, for k 6= −n, 0, n, splitting p probability
to z + (k+1)α
and z − (k−1)α
will increase the variance by p2 ( αn )2 .
from z + kα
n
n
n
Therefore, for positive even i we have
1 − 2L α 2
σ 2 (Zi ) − σ 2 (Zi−2 ) >
2n − 1 n
If L < 21 , then after enough steps the variance of Zi will exceed α2 , a contradiction.
That we can do part 1 of the theorem for all lotteries Y follows by the
fact that a countable set of countable sequences is countable.
Finally, the following lemma proves part 2 of the theorem.
Lemma 6 Any sequence of left symmetric split starting at δµ converges (in
the L1 topology) to a skewed to the left distribution with expected value µ.
Proof: That such sequences converge follows from the fact that a symmetric
split will increase the variance of the distribution, but as all distributions
are over the bounded [x, x] segment of ℜ, the variances of the distributions
increase to a limit. Replacing (x, p) with (x − α, p2 ; x + α, p2 ) increases the
variance of the distribution by
p
p
(x − α − µ)2 + (x + α − µ)2 − p(x − µ)2 = pα2
2
2
and therefore the distance between two successive distributions in the sequences in bounded by x − x times the change in the variance. The sum of
the changes in the variances is bounded, as is therefore the sum of distances
between successive distributions, hence Cauchy criterion is satisfied and the
sequence converges.
Next we prove that the limit is a skewed to the left distribution with
expected value µ. Let F be the distribution of X = (x1 , p1 ; . . . ; xn , pn ) with
expected value µ be skewed to the left. Suppose wlg that x1 6 µ, and break
it symmetrically to obtain X ′ = (x1 − α, p21 ; x1 + α, p21 ; x2 , p2 ; . . . ; xn , pn ) with
the distribution F ′ . Note that E[X ′ ] = µ. Consider the following two cases.
Case 1: x1 + α 6 µ. Then for all δ, η2 (F, δ) = η2 (F ′ , δ). For δ such that
µ − δ 6 x1 − α or such that x1 + α 6 µ − δ, η1 (F ′ , δ) = η1 (F, δ) > η2 (F, δ) =
η2 (F ′ , δ). For δ such that x1 − α < µ − δ 6 x1 , η1 (F ′ , δ) = η1 (F, δ) + [(µ −
32
δ) − (x1 − α)] p21 > η1 (F, δ) > η2 (F, δ) = η2 (F, δ). Finally, for δ such that
x1 < µ − δ < x1 + α ( 6 µ), η1 (F ′ , δ) = η1 (F, δ) + [(x1 + α) − (µ − δ)] p21 >
η1 (F, δ) > η2 (F, δ) = η2 (F ′ , δ).
Case 2: x1 + α > µ. Then for all δ such that µ + δ > x1 + α, η2 (F, δ) =
η2 (F ′ , δ). For δ such that µ − δ 6 x1 − α, η1 (F ′ , δ) = η1 (F, δ) > η2 (F, δ) =
η2 (F ′ , δ). For δ such that x1 −α < µ−δ 6 x1 , η1 (F ′ , δ) = η1 (F, δ)+[(µ−δ)−
(x1 −α)] p21 > η2 (F, δ)+[(µ−δ)−(x1 −α)] p21 > η2 (F, δ)+max{0, (x1 +α)−(µ+
δ)} p21 = η2 (F ′ , δ). Finally, for δ such that µ − δ > x1 , η1 (F ′ , δ) = η1 (F, δ) +
[(x1 + α) − (µ − δ)] p21 > η2 (F, δ) + max{0, (x1 + α) − (µ + δ)} p21 = η2 (F ′ , δ).
If Xn → Y , all have the same expected value and for all n, Xn is skewed
to the left, then so is Y .
Proof of Theorem 3: We first note the following immediate corollary of
Theorem 1.
Corollary 1 Under the assumptions of Theorem 1, the decision maker rejects skewed-to-the-left splits of outcomes below the mean.
In proving Theorem 3, we will use the following rules:
a. Rejection of left symmetric splits (Theorem 1). For example, (8, 21 ; 4, 12 ) ≻a
(8, 12 ; 6, 14 ; 2, 14 ).
b. Rejection of larger left-symmetric splits (Definition 8). For example,
(8, 21 ; 6, 14 ; 2, 14 ) ≻b (8, 12 ; 7, 14 ; 1, 14 ).
c. Rejection of skewed-to-the-left splits of outcomes below the mean (Corollary 1). For example, (10, 12 ; 6, 12 ) ≻c (10, 21 ; 7, 13 ; 4, 16 ).
The proof is divided into two parts.
Part 1. Suppose first that q > p. Since p > 12 ,
1 − (1 − q + p) 6 1 − q + p −
1−q
1−p
Hence
(1 − q + p, 1) c
1−q
1, p;
,1 − p
1−p
33
= X1
Case 1.1 If q 6
2p
,
1+p
then
1−
1−q
1−q
6
− (1 − q)
1−p
1−p
In which case
X1 =
1−q
1, p;
,1 − p
1−p
Fig. 3 depicts this case for p =
2
3
c
q−p
p
1, ; 1 − q,
q
q
= Y1
and q = 43 .
1−p
X
Y
1− p
q
1−q
1−p
1−q
0
1−q+p
1
Figure 3: Case 1.1
Case 1.2 If q >
2p
1+p
but q 6 2p − p2 , then
1 − (1 + p − q) 6
1−q
− (1 − q)
1−p
And
p
(1 − p) − 1 −
q
<1−
p
q
Now
1−q
, 1 − p a
X1 = 1, p;
1−p
p
p
2p
Z1 := 1, p; 1 − q + p, − p; 1 − p, − p; 1 − q, 1 + p −
q
q
q
34
Explanation: Z1 is a mean preserving spread of X1 and can therefore be
obtained from the latter by a sequence of symmetric splits. Since X1 and Z1
differ only for outcomes below the average 1 + p − q, all these splits are left
symmetric splits. Finally,
q 2p
2p
1, p; 1 − ,
a Z1 b Y 1
− 2p; 1 − q, 1 + p −
2 q
q
Fig. 4 depicts this case for p =
heavy line.
2
3
and q = 65 . Lottery Z1 is depicted by the
1−p
X
Y
1− p
q
Y
Z
1+p− 2p
q
0
1−q
1−q
1−p
1−p
1− q
2
1−q+p
1
Figure 4: Case 1.2
Case 1.3 Suppose now that q >
2p
1+p
1 − (1 + p − q) >
And
and q > 2p − p2 . Then
1−q
− (1 − q)
1−p
p
(1 − p) − 1 −
q
<1−
p
q
Similarly to the analysis of case 2,
q 2p
2p
′
X1 a Z1 := 1, p; 1 − ,
− 2p; 1 − q, 1 + p −
2 q
q
Observe that Z1′ is is mean preserving spread of X1 and that the two differ
only below the average 1−q +p. The case follows by Z1′ a Y1 . Fig. 5 depicts
11
. Lottery Z1′ is depicted by the heavy line.
this case for p = 23 and q = 12
35
X
1−p
1−p/q
1+p−2p/q
Y
Z′
Y
1−q
1−p
1−q
0
1− q
2
1−q+p
1
Figure 5: Case 1.3
Part 2: Here we deal with the case q < p and show that X2 Y2 .
Case 2.1 If q >
p
2−p
then 1 −
q
p
6
Case 2.2 Suppose that p2 6 q <
1 − (1 − p + q) 6
q
p
− q hence X2 c Y2 .
p
.
2−p
Then
q
2(p − q)
− q and p 6
p
1−q
Therefore
X2 a
p − 2q + pq
p−q
; q,
1, 1 − p; p, p −
1−q
1−q
a Y2
Case 2.3 Suppose that 2p − 1 6 q 6 p2 . Then
1 − (1 − p + q) >
q
−q
p
(11)
And the outcome 1 − ( pq − q) is to the right of the average 1 − p + q. Also,
p6
2(p − q)
⇐⇒ p − 2q + pq > 0
1−q
The expression p−2q+pq is decreasing in q, and at q = p2 it is p(1−2p+p2 ) =
p(1 − p)2 > 0. Therefore
(m + s)q(1 − p)
p−q
p−q
=
= (m + s) p −
1−q
1−q
1−q
36
Where m is a positive integer and 0 6 s < 1. We now move from X2 to Y2
in m + 1 utility-reducing steps (If s = 0, then there are only m steps).
Step 0: Move from X2 to
sq(1 − p) q mq(1 − p) q
q(1 − p
q
0
Z2 = q,
; ,
; +s
−q ,
; 1, 1 − p
1−q
p
1−q
p
p
1−q
This is a left split to the left of the average, hence by Theorem 1 X2 c Z20 .
If s = 0 denote Z20 = X2 .
Steps 1, . . . , m: For i = 1, . . . , m, let
(i + s)q(1 − p) q (m − i)q(1 − p)
i
; ,
;
Z2 =
q,
1−q
p
1−q
q
q(1 − p
q
+ (i + s)
−q ,
; 1, 1 − p
p
p
1−q
Observe that Z2i is obtained from Z2i−1 by a continuation of a symmetric split
around
q
1 q
+ (i + s)
−q
6
di : =
2 p
p
1 q
q
1+q
+ 1− +q
=
6 1 − p + q ⇐⇒
2 p
p
2
2p 6 1 + q
We assumed q > 2p−1, hence these splits are all symmetric around outcomes
(weakly) below the average 1 − p + q. Therefore
X2 c Z20 b Z21 b . . . b Z2m = Y2
Lottery Z2m−1 is depicted by the heavy line in Fig. 6 which is drawn for
the case p = 32 and q = 13 . Observe that in this case s = 0.
Proof of Proposition 2: Suppose that for q sufficiently close to zero,
h1, q; 0, 1 − qi ≻ hq, 1i. These preferences imply that the derivative of h(q) :=
V (1, q; 0, 1 − q) − V (q, 1) at q = 0 is positive. We obtain:
h′ (q) = V2 (1, q; 0, 1 − q) − V4 (1, q; 0, 1 − q) − V1 (q, 1; 0, 0) =⇒
h′ (0) = V2 (1, 0; 0, 1) − V4 (1, 0; 0, 1) − V1 (0, 1; 0, 0) > 0
(12)
37
X
1−p
Y
Z
Y
m−1
p−q
1−q
X
q
0
q
p
1−p+q
q
1− p
+q
1
Figure 6: Case 2.3
; q, p−q
i,
Consider now the lotteries X2 = h1, 1 − p; pq , pi and Y2 = h1, 1−p
1−q
1−q
and let p = 1 − kq. We want to check conditions under which for sufficiently
small q, g(q) := V (Y2 ) − V (X2 ) > 0. In other words, we want to show that
the first non-zero derivative with respect to q of g at q = 0 is positive. Let
kq
q
A = (1, 1−q
; q, 1−(k+1)q
) and B = (1, kq; 1−kq
, 1 − kq). We get
1−q
g(q) = V (A) − V (B)
k
g ′ (q) = (1−q)
2 [V2 (A) − V4 (A)] + V3 (A) −
k[V2 (B) − V4 (B)] −
1
V (B)
(1−kq)2 3
=⇒
g ′ (0) = k[V2 (A) − V4 (A)] + V3 (A) − k[V2 (B) − V4 (B)] − V3 (B)
At q = 0, V2 (A) − V4 (A) and V2 (B) − V4 (B) describe the same effect: the
change in the value of the lottery (1, r; 0, 1 − r) as the probability r starts to
go up from zero. Likewise, at q = 0, V3 (A) and V3 (B) describe the change
in the value of the lottery (1, 0; s, 1) as the outcome s starts to go up from
zero. It thus follows that g ′ (0) = 0.
g ′′ (q) =
2k
k2
[V (A) − V4 (A)] + (1−q)
4 [V22 (A) − 2V24 (A) + V44 (A)]
(1−q)3 2
2k
[V (A) − V43 (A)] + V33 (A) −
(1−q)2 23
2k
k 2 [V22 (B) − 2V24 (B) + V44 (B)] − (1−kq)
2 [V23 (B) − V43 (B)]
1
− (1−kq)
4 V33 (B) =⇒
V4 (A)] + k 2 [V22 (A) −
2k
V (B)
(1−kq)3 3
g ′′ (0) = 2k[V2 (A) −
2V24 (A) + V44 (A)] +
2k[V23 (A) − V43 (A)] + V33 (A) −
38
+
−
k 2 [V22 (B) − 2V24 (B) + V44 (B)] − 2k[V23 (B) − V43 (B)] −
2kV3 (B) − V33 (B)
Similarly to the previous analysis, at q = 0, Vij (A) = Vij (B) for all i and j,
hence
g ′′ (0) = 2k[V2 (A) − V4 (A) − V3 (B)]
At q = 0, A = B = (1, 0; 0, 1) ∼ (0, 1; 1, 0). Moreover, for all a, (a, 1; 0, 0) ∼
(a, 1; 1, 0). Therefore, at q = 0, V3 (B) = V1 (0, 1; 0, 0). It thus follows that
g ′′ (0) = 2kh′ (0) > 0.
References
[1] Abdellaoui, M., P. Klibanoff, and L. Placido, 2013. “Experiments on
compound risk in relation to simple risk and to ambiguity,” mimeo.
[2] Abdulkadiroğlu, A. and T. Sönmez, 1998. “Random serial dictatorship
and the core from random endowments in house allocation problems,”
Econometrica 66:689–701.
[3] Abdulkadiroğlu, A. and T. Sönmez, 2013. “Matching markets: Theory
and practice,” Advances in Economics and Econometrics, D. Acemoglu,
M. Arello, and E. Dekel (eds), Cambridge: CUP.
[4] Becker, S.W. and F.O. Brownson, 1964: “What price ambiguity? Or
the role of ambiguity in decision making,” Journal of Political Economy
72:62–73.
[5] Bernasconi, M. and G. Loomes, 1992. “Failures of the reduction principle
in an Ellsberg-type problem,” Theory and Decision 32:77–100.
[6] Boiney, L.G., 1993. “The effects of skewed probability on decision making under ambiguity,” Organizional Behavior and Human Decisions Processes 56:134–148.
[7] Budish, E., Y.-K. Che, F. Kojima, and P. Milgrom, 2013. “Designing
random allocation mechanisms: Theory and applications,” American
Economic Review 103:585–623.
39
[8] Camerer, C. F., 1995.“Individual decision making,” in Handbook of Experimental Economics, ed. by A.R.J. Kagel, Princeton University Press,
Princeton, vol. 2.
[9] Cerreia Vioglio, S., D. Dillenberger, and P. Ortoleva, 2014. “Cautious
expected utility and the certainty effect,” Econometrica, forthcoming.
[10] Chew, S.H., L.G. Epstein, and U. Segal, 1991. “Mixture symmetry and
quadratic utility,” Econometrica 59:139–163.
[11] Chew S.H., B. Miao, and S. Zhong, 2013. “Partial ambiguity,” working
paper.
[12] Chew, S.H. and W.S. Waller, 1986. “Empirical tests of weighted utility
theory,” Journal of Mathematical Psychology 30:55–72.
[13] Conlisk, J., 1989. “Three variants on the Allais example,” American
Economic Review 79:392–407.
[14] Coombs, C. and L. Huang, 1976. “Tests of the betweenness property of
expected utility,” Journal of Mathematical Psychology 13:323-337.
[15] Dekel, E., 1989. “Asset demands without the independence Axiom,”
Econometrica 57:163–69.
[16] Dillenberger, D., 2010. “Preferences for one-shot resolution of uncertainty and Allais-type behavior,” Econometrica 78:1973–2004.
[17] Dwenger, N., D. Kubler, and G. Weizsacker, 2013. “Flipping a coin:
Theory and evidence,” Mimeo, Humboldt-Universität zu Berlin.
[18] Ellsberg, D., 1961. “Risk, ambiguity, and the Savage axioms,” Quarterly
Journal of Economics 75:643–669.
[19] Ghirardato, P., Maccheroni, F., and Marinacci, M., 2004. “Differentiating ambiguity and ambiguity attitude,” Journal of Economic Theory
118:133–73.
[20] Gilboa, I. and D. Schmeidler, 1989. “Maxmin expected utility with a
non-unique prior,” Journal of Mathematical Economics 18:141–153.
40
[21] Halevy, Y., 2007. “Ellsberg revisited: An experimental study,” Econometrica 75:503–536.
[22] Harrison, G. W., J. Martinez-Correa, and J.T. Swarthout, 2012. “Reduction of compound lotteries with objective probabilities: Theory and
evidence,” working paper.
[23] Hylland, A., and Zeckhauser, R., 1979. “The efficient allocation of individuals to positions,” Journal of Political Economy 87:293–314.
[24] Kahn, B.E. and R.K. Sarin, 1988. “Modeling ambiguity in decisions
under uncertainty,” Journal of Consumer Research 15:265–272.
[25] Kahneman, D. and A. Tversky, 1979. “Prospect theory: An analysis of
decision under risk,” Econometrica 47:263–291.
[26] Kreps, D.M. and E.L. Porteus, 1978. “Temporal resolution of uncertainty and dynamic choice theory,” Econometrica 46:185–200.
[27] Klibanoff, P., M. Marinacci, and S. Mukerji, 2005. “A smooth model of
decision making under ambiguity,” Econometrica 73:1849–1892.
[28] Machina, M.J., 1982. “ ‘Expected utility’ analysis without the independence axiom,” Econometrica 50:277–323.
[29] Maccheroni, F., M. Marinacci, and A. Rustichini, 2006. “ Ambiguity
aversion, robustness, and the variational representation of preferences,”
Econometrica 74:1447–1498
[30] Masatlioglu, Y., Y. Orhun, and C. Raymond, 2015. “Skewness and intrinsic preferences for information, Mimeo.
[31] Menezes, C.F., C. Geiss, and J. Tressler, 1980. “Increasing downside
risk,” American Economic Review 70:921–932.
[32] Miao, B., and S. Zhong, 2012. “An experimental study of attitude towards two-stage lottery,” working paper.
[33] Pathak, P.A. and J. Sethuraman, 2011. “Lotteries in student assignment:
An equivalence result,” Theoretical Economics 6:1–17.
41
[34] Quiggin, J., 1982. “A theory of anticipated utility,” Journal of Economic
Behavior and Organization 3:323–343.
[35] Rothschild, M. and J.E. Stiglitz, 1970. “Increasing risk: I. A definition,”
Journal of Economic Theory 2:225–243.
[36] Schmeidler, D., 1989. “Subjective probability and expected utility without additivity,” Econometrica 57:571–587.
[37] Segal, U., 1987. “The Ellsberg paradox and risk aversion: An anticipated
utility approach,” International Economic Review 28:175–202.
[38] Segal, U., 1990. “Two-stage lotteries without the reduction axiom,”
Econometrica 58:349–377.
[39] Segal, U. and A. Spivak, 1990. “first-order versus second-order risk aversion,” Journal of Economic Theory 51:111–125.
[40] Sopher, B. and J. M. Narramore, 2000. “Stochastic choice and consistency in decision making under risk: An experimental study,” Theory
and Decision 48:323-350.
[41] van de Kuilen, G. and P.P. Wakker, 2011. “The midweight method
to measure attitudes toward risk and ambiguity,” Management Science
57:582–598.
[42] Viscusi, W.K. and H. Chesson, 1999: “Hopes and fears: The confliction
effects of risk ambiguity,” Theory and Decision 47:153–178.
42
Appendix A: Example 1
A quadratic
utility (Chew, Epstein, and Segal [10]) functional is given by
V (p) = x y px py θ(x, y), where θ is symmetric. Following [10, Example 5
, where v and w are positive functions,
(p. 145)], If θ(x, y) = v(x)w(y)+v(y)w(x)
2
then V (p) = E[v(p)] × E[w(p)]. This is the form of V we analyze below.
The function V is the product of two positive linear functions of the
probabilities, hence quasi concave. To see why, observe that ln V (p) =
ln E[v(p)] + ln E[w(p)]. The sum of concave functions is concave, hence quasi
concave, and any monotone nondecreasing transformation of a quasi concave
function is quasi concave.
Direct calculations show that
the local utility function of any quadratic
utility is given by uF (x) = 2 θ(x, y)dF (y). Since we are only interested in
the behavior of the function in lotteries of the form δy := (y, 1), we have
uδy (x) = 2θ(x, y) = v(x)w(y) + v(y)w(x)
Take v(x) = x and let w be any increasing, concave, and differential
function such that w(0) = 0. We now show that V satisfies Weak Hypothesis
II. That is, we show that
RA := −
uδy (x)
uδy (x)
=−
yw (x)
w(y) + yw (x)
is an increasing function of y. We have
yw (x)
∂
> 0 ⇐⇒
−
∂y w(y) + yw (x)
w (x)(w(y) + yw (x)) < (w (y) + w (x))yw (x) ⇐⇒
w(y) > w (y)y ⇐⇒
w(y)/y > w (y)
which holds since w is concave.
Next we analyze the functional form V (p1 , q1 ; . . . ; pn , qn ) = E[w(cp )] ×
ζ
, cp = βp + (1 − β)pκ , ζ = 1.024,16 κ = 1.1, and
E[cp ] where w(x) = ζx−x
ζ−1
16
since ζ > 1, we have that w (x) =
is increasing and concave.
ζ−ζxζ−1
ζ−1
43
> 0 and w (x) =
(ζ−1)ζxζ−2
ζ−1
< 0, hence w
β = 0.15. Since all the inequalities below are strict, there is an open set of
parameters for which they are satisfied as well. Observe that
ζ [βp + (1 − β)pκ ] − [βp + (1 − β)pκ ]ζ
w(cp ) =
ζ −1
We show first that this functional rejects all symmetric noise. For any
0 < p < 1 and ε min{p, 1 − p}, let
f (ε, p) := [w(cp+ε ) + w(cp−ε )] × [cp+ε + cp−ε ]
Rejection of symmetric noise requires that f (0, p) − f (ε, p) > 0 for all p ∈
(0, 1) and ε ∈ (0, min{p, 1 − p}). Numerical calculations show that this is
indeed the case. See graph below.
Figure A1: Rejection of symmetric noise
Using the same functional as above, we now show that for every p > 0
there exists a sufficiently small q > 0 such that p, q; 0, 1 − q pq, 1, that
is, the decision maker always accepts some positively skewed noise.
For q = 0, V (cpq , 1) − V (cp , q; 0, 1 − q) = 0. We show that for every p < 1,
the first non-zero derivative of this expression with respect to q at q = 0 is
negative. We get
(ζ − 1)V (cpq , 1) = (ζ − 1)w(cpq )cpq =
ζ [βpq + (1 − β)pκ q κ ] − [βpq + (1 − β)pκ q κ ]ζ × [βpq + (1 − β)pκ q κ ]
44
Differentiate with respect to q to obtain
ζ βp + κ(1 − β)pκ q κ−1 − ζ [βpq + (1 − β)pκ q κ ]ζ−1 βp + κ(1 − β)pκ q κ−1 ×
ζ [βpq + (1 − β)pκ q κ ] − [βpq + (1 − β)pκ q κ ]ζ
[βpq + (1 − β)pκ q κ ] +
× βp + κ(1 − β)pκ q κ−1
At q = 0, this expression equals 0. Differentiate again with respect to q to
obtain
2
ζ κ(κ − 1)(1 − β)pκ q κ−2 − (ζ − 1) [βpq + (1 − β)pκ q κ ]ζ−2 βp + κ(1 − β)pκ q κ−1 −
[βpq + (1 − β)pκ q κ ]ζ−1 κ(κ − 1)(1 − β)pκ q κ−2 × [βpq + (1 − β)pκ q κ ] +
κ κ−1
κ κ ζ−1
κ κ−1
− [βpq + (1 − β)p q ]
βp + κ(1 − β)p q
2ζ βp + κ(1 − β)p q
×
βp + κ(1 − β)pκ q κ−1 +
ζ [βpq + (1 − β)pκ q κ ] − [βpq + (1 − β)pκ q κ ]ζ × κ(κ − 1)(1 − β)pκ q κ−2
Observe that
2
ζ κ(κ − 1)(1 − β)pκ q κ−2 − (ζ − 1) [βpq + (1 − β)pκ q κ ]ζ−2 βp + κ(1 − β)pκ q κ−1 −
[βpq + (1 − β)pκ q κ ]ζ−1 κ(κ − 1)(1 − β)pκ q κ−2 × [βpq + (1 − β)pκ q κ ] =
2
ζ κ(κ − 1)(1 − β)pκ q κ−1 − (ζ − 1)q [βpq + (1 − β)pκ q κ ]ζ−2 βp + κ(1 − β)pκ q κ−1 −
κ κ ζ−1
κ κ−1
[βpq + (1 − β)p q ] κ(κ − 1)(1 − β)p q
× βp + (1 − β)pκ q κ−1
This expression converges to zero with q. This is obvious for ζ 2. If
2 > ζ > 1, then notice that by l’Hospital’s rule
lim
q→0
q
[βpq + (1 − β)pκ q κ ]2−ζ
[βpq + (1 − β)pκ q κ ]ζ−1
=0
q→0 (2 − ζ) [βp + κ(1 − β)pκ q κ−1 ]
= lim
Also, as q → 0, the limit of the expression
2ζ βp + κ(1 − β)pκ q κ−1 − [βpq + (1 − β)pκ q κ ]ζ−1 βp + κ(1 − β)pκ q κ−1 ×
βp + κ(1 − β)pκ q κ−1
45
is 2ζβ 2 p2 . Finaly,
κ κ
κ κ ζ
ζ [βpq + (1 − β)p q ] − [βpq + (1 − β)p q ] × κ(κ − 1)(1 − β)pκ q κ−2 =
ζ 1− ζ1
κ κ−1
κ κ− ζ1
ζ βp + (1 − β)p q
− βpq
× κ(κ − 1)(1 − β)pκ q κ−1
+ (1 − β)p q
As ζ, κ > 1, this expression goes to zero with q.
On the other hand, (ζ − 1)V (cp , q; 0, 1 − q) equals
q 2 ζ [βp + (1 − β)pκ ] − [βp + (1 − β)pκ ]ζ × [βp + (1 − β)pκ ]
Its first order derivative with respect to q at q = 0 is zero, while the second
derivative at this point equals
2 ζ [βp + (1 − β)pκ ] − [βp + (1 − β)pκ ]ζ × [βp + (1 − β)pκ ]
We therefore get that the first order derivative of V (cpq ), 1) − V (cp , q; 0, 1 − q)
at q = 0 is zero, and that
∂2
[V (cpq ), 1) − V (cp , q; 0, 1 − q)] = g(p; β, ζ, κ) :=
q→0 ∂q 2
2 2
κ
κ ζ
2ζβ p − 2 ζ [βp + (1 − β)p ] − [βp + (1 − β)p ] × [βp + (1 − β)pκ ]
(ζ − 1) lim
The graph below shows g(p; β, ζ, κ) for β = 0.15, κ = 1.1, and ζ = 1.024.
Note that for these values g(p; β, ζ, κ) < 0 for all p ∈ (0, 1), which means
that for q > 0 small enough, the positively skewed noise p, q; 0, 1 − q is
accepted.
−3
0
x 10
−1
−2
−3
g(p)
−4
−5
−6
−7
−8
−9
0
0.1
0.2
0.3
0.4
0.5
p
0.6
0.7
0.8
0.9
Figure A2: g(p; 0.15, 1.024, 1.1)
46
1
Finally, to illustrate why Proposition 2 and the predictions of Theorem 3
are not contradictory, we provide numerical results showing that TC is always
preferred to SD if q 2p − 1. Recall that w and cp are bounded in [0, 1].
The following graphs show the combinations of p and q for which the value
of TC exceeds the value of SD by 0.01 and 0.001. In both pictures, the left
panel is for the case q > p and the right panel is for the case q < p. In both
cases, all pairs (p, q) such that q 2p − 1 satisfy the requirements.
TC-SD>=0.01
1
TC-SD>=0.01
0.7
0.95
0.6
0.9
0.5
0.85
0.4
0.75
q
q
0.8
0.3
0.7
0.65
0.2
0.6
0.1
0.55
0.5
0.5
0.55
0.6
0.65
0.7
0.75
p
0.8
0.85
0.9
0.95
0
0.5
1
0.55
0.6
0.65
0.7
0.75
p
0.8
0.85
0.9
Figure A3: V (T C) − V (SD) > 0.01
TC-SD>=0.001
0.9
0.8
0.85
0.7
0.8
0.6
0.75
0.5
0.7
0.4
0.65
0.3
0.6
0.2
0.55
0.5
0.5
TC-SD>=0.001
1
0.9
q
q
1
0.95
0.1
0.55
0.6
0.65
0.7
0.75
p
0.8
0.85
0.9
0.95
0
0.5
1
0.55
0.6
0.65
0.7
0.75
p
0.8
Figure A4: V (T C) − V (SD) > 0.001
47
0.85
0.9
0.95
1
Download